Our client is a leading manufacturer of industrial products. The client’s Fluids division provides an integrated solution where strands of polymer are passed to the rotor which cuts the polymer in the desired configuration. The continuous usage of the rotor makes its blade dull leading to unusable pellets of wrong dimensions. This resulted in the wastage of resources and impacted the client’s revenue.
The main objective of this engagement was to identify the optimal number of sensors and the best positions to mount them on machines in order to detect the transition of rotor blade states from sharp to dull.
- Huge volume of data from multiple sensors
- Difficulty in attaining high model accuracy as sensor data is time-based and requires multiple transformations
- Need for a lightweight solution as the client wanted to deploy the solution on edge devices
- Collated input datasets from sensor data and rotor configuration data
- Collected blade classification data at a frequency of 200Hz for 2 minutes for each Rotor configuration
- Performed data cleansing and Feature Engineering to prepare a model-ready dataset
- Combined data across individual sensor positions from all machines and used them to train the model
- Retrained the Model using important features to make it lightweight and improve inference time for edge deployment
- Created separate models for each sensor position to identify which position is ideal to detect blade wear & tear
- Applied multiple independent approaches (FFT, Wavelets, PSD) to achieve high accuracy
- Achieved a 66% reduction in data and cost of sensors
- Achieved an accuracy of ~100% and were able to clearly classify blades from the identified optimal sensor positions
- Deployed solution on edge by developing a lightweight solution through computationally inexpensive features without compromising on accuracy for real-time inferencing
- Sensor deployed on optimal position will be used for blade classification in the future